Why we trust machines

Cover of Why we trust machines, Frontier Report No. 01

Frontier Report No. 01 · 21 pages · 17 sources

Should you trust advice from AI?

Sometimes, and the skill is knowing when. In a field experiment with 758 consultants, AI raised speed and quality on tasks within its competence and lowered accuracy on a task beyond it, and people could not easily tell which was which. The goal is trust that rises and falls with how reliable a system is for the task in hand.

What the report found

  1. Use has outrun trust. In a 2025 survey of more than 48,000 people in 47 countries, two thirds used AI regularly, but fewer than half said they were willing to trust it. Two thirds also said they relied on its output without checking it.
  2. People both shun and over-trust machines. Older research shows people abandon an algorithm after seeing it err, yet lay people often prefer algorithmic advice to human advice. Which wins depends on the task, the person and whether a mistake has been seen.
  3. AI help is uneven. In a field experiment with 758 consultants, AI raised speed and quality on tasks within its competence and lowered accuracy on a task beyond it. Users could not easily tell which was which.
  4. Human and machine together are not automatically better. A review of 106 experiments found that, on average, combinations did worse than the better of the two working alone, especially for decisions.
  5. The real question is calibration. The goal is not more or less trust in machines, but trust that rises and falls with how reliable a system is for the task in hand. How to achieve that is largely untested.

What is inside

  1. What is new
  2. What older research tells us
  3. What early studies suggest
  4. What nobody knows yet
  5. Questions worth watching

Plus an evidence maturity map, a scenario grid, signals to watch, a glossary and the full list of sources.

Who it is for

For anyone who builds, buys or relies on advice from software.

How we work

Frontier Reports cover behaviour that is moving faster than the research. Every claim is sorted into Established, Emerging or Untested, so an early finding is never read as a settled one. Where researchers disagree, as they do on whether people distrust or over-trust algorithms, both sides are set out before we say where we stand.

Cite this report

The Why We Project (2026) Why we trust machines: Advice from software. Frontier Report No. 01. Dubai: Why We Tech Ltd. Available at: thewhyweproject.com